1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */ 2 /************************************************************************** 3 * 4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 5 * All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a 8 * copy of this software and associated documentation files (the 9 * "Software"), to deal in the Software without restriction, including 10 * without limitation the rights to use, copy, modify, merge, publish, 11 * distribute, sub license, and/or sell copies of the Software, and to 12 * permit persons to whom the Software is furnished to do so, subject to 13 * the following conditions: 14 * 15 * The above copyright notice and this permission notice (including the 16 * next paragraph) shall be included in all copies or substantial portions 17 * of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 25 * USE OR OTHER DEALINGS IN THE SOFTWARE. 26 * 27 **************************************************************************/ 28 /* 29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 30 */ 31 32 #define pr_fmt(fmt) "[TTM] " fmt 33 34 #include <drm/ttm/ttm_module.h> 35 #include <drm/ttm/ttm_bo_driver.h> 36 #include <drm/ttm/ttm_placement.h> 37 #include <linux/jiffies.h> 38 #include <linux/slab.h> 39 #include <linux/sched.h> 40 #include <linux/mm.h> 41 #include <linux/file.h> 42 #include <linux/module.h> 43 #include <linux/atomic.h> 44 #include <linux/reservation.h> 45 46 static void ttm_bo_global_kobj_release(struct kobject *kobj); 47 48 static struct attribute ttm_bo_count = { 49 .name = "bo_count", 50 .mode = S_IRUGO 51 }; 52 53 /* default destructor */ 54 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo) 55 { 56 kfree(bo); 57 } 58 59 static inline int ttm_mem_type_from_place(const struct ttm_place *place, 60 uint32_t *mem_type) 61 { 62 int pos; 63 64 pos = ffs(place->flags & TTM_PL_MASK_MEM); 65 if (unlikely(!pos)) 66 return -EINVAL; 67 68 *mem_type = pos - 1; 69 return 0; 70 } 71 72 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 73 { 74 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 75 struct drm_printer p = drm_debug_printer(TTM_PFX); 76 77 pr_err(" has_type: %d\n", man->has_type); 78 pr_err(" use_type: %d\n", man->use_type); 79 pr_err(" flags: 0x%08X\n", man->flags); 80 pr_err(" gpu_offset: 0x%08llX\n", man->gpu_offset); 81 pr_err(" size: %llu\n", man->size); 82 pr_err(" available_caching: 0x%08X\n", man->available_caching); 83 pr_err(" default_caching: 0x%08X\n", man->default_caching); 84 if (mem_type != TTM_PL_SYSTEM) 85 (*man->func->debug)(man, &p); 86 } 87 88 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 89 struct ttm_placement *placement) 90 { 91 int i, ret, mem_type; 92 93 pr_err("No space for %p (%lu pages, %luK, %luM)\n", 94 bo, bo->mem.num_pages, bo->mem.size >> 10, 95 bo->mem.size >> 20); 96 for (i = 0; i < placement->num_placement; i++) { 97 ret = ttm_mem_type_from_place(&placement->placement[i], 98 &mem_type); 99 if (ret) 100 return; 101 pr_err(" placement[%d]=0x%08X (%d)\n", 102 i, placement->placement[i].flags, mem_type); 103 ttm_mem_type_debug(bo->bdev, mem_type); 104 } 105 } 106 107 static ssize_t ttm_bo_global_show(struct kobject *kobj, 108 struct attribute *attr, 109 char *buffer) 110 { 111 struct ttm_bo_global *glob = 112 container_of(kobj, struct ttm_bo_global, kobj); 113 114 return snprintf(buffer, PAGE_SIZE, "%d\n", 115 atomic_read(&glob->bo_count)); 116 } 117 118 static struct attribute *ttm_bo_global_attrs[] = { 119 &ttm_bo_count, 120 NULL 121 }; 122 123 static const struct sysfs_ops ttm_bo_global_ops = { 124 .show = &ttm_bo_global_show 125 }; 126 127 static struct kobj_type ttm_bo_glob_kobj_type = { 128 .release = &ttm_bo_global_kobj_release, 129 .sysfs_ops = &ttm_bo_global_ops, 130 .default_attrs = ttm_bo_global_attrs 131 }; 132 133 134 static inline uint32_t ttm_bo_type_flags(unsigned type) 135 { 136 return 1 << (type); 137 } 138 139 static void ttm_bo_release_list(struct kref *list_kref) 140 { 141 struct ttm_buffer_object *bo = 142 container_of(list_kref, struct ttm_buffer_object, list_kref); 143 struct ttm_bo_device *bdev = bo->bdev; 144 size_t acc_size = bo->acc_size; 145 146 BUG_ON(kref_read(&bo->list_kref)); 147 BUG_ON(kref_read(&bo->kref)); 148 BUG_ON(atomic_read(&bo->cpu_writers)); 149 BUG_ON(bo->mem.mm_node != NULL); 150 BUG_ON(!list_empty(&bo->lru)); 151 BUG_ON(!list_empty(&bo->ddestroy)); 152 ttm_tt_destroy(bo->ttm); 153 atomic_dec(&bo->bdev->glob->bo_count); 154 dma_fence_put(bo->moving); 155 reservation_object_fini(&bo->ttm_resv); 156 mutex_destroy(&bo->wu_mutex); 157 bo->destroy(bo); 158 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 159 } 160 161 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 162 { 163 struct ttm_bo_device *bdev = bo->bdev; 164 struct ttm_mem_type_manager *man; 165 166 reservation_object_assert_held(bo->resv); 167 168 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 169 BUG_ON(!list_empty(&bo->lru)); 170 171 man = &bdev->man[bo->mem.mem_type]; 172 list_add_tail(&bo->lru, &man->lru[bo->priority]); 173 kref_get(&bo->list_kref); 174 175 if (bo->ttm && !(bo->ttm->page_flags & 176 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) { 177 list_add_tail(&bo->swap, 178 &bdev->glob->swap_lru[bo->priority]); 179 kref_get(&bo->list_kref); 180 } 181 } 182 } 183 EXPORT_SYMBOL(ttm_bo_add_to_lru); 184 185 static void ttm_bo_ref_bug(struct kref *list_kref) 186 { 187 BUG(); 188 } 189 190 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 191 { 192 if (!list_empty(&bo->swap)) { 193 list_del_init(&bo->swap); 194 kref_put(&bo->list_kref, ttm_bo_ref_bug); 195 } 196 if (!list_empty(&bo->lru)) { 197 list_del_init(&bo->lru); 198 kref_put(&bo->list_kref, ttm_bo_ref_bug); 199 } 200 201 /* 202 * TODO: Add a driver hook to delete from 203 * driver-specific LRU's here. 204 */ 205 } 206 207 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo) 208 { 209 struct ttm_bo_global *glob = bo->bdev->glob; 210 211 spin_lock(&glob->lru_lock); 212 ttm_bo_del_from_lru(bo); 213 spin_unlock(&glob->lru_lock); 214 } 215 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru); 216 217 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo) 218 { 219 reservation_object_assert_held(bo->resv); 220 221 ttm_bo_del_from_lru(bo); 222 ttm_bo_add_to_lru(bo); 223 } 224 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail); 225 226 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 227 struct ttm_mem_reg *mem, bool evict, 228 struct ttm_operation_ctx *ctx) 229 { 230 struct ttm_bo_device *bdev = bo->bdev; 231 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 232 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 233 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 234 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 235 int ret = 0; 236 237 if (old_is_pci || new_is_pci || 238 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 239 ret = ttm_mem_io_lock(old_man, true); 240 if (unlikely(ret != 0)) 241 goto out_err; 242 ttm_bo_unmap_virtual_locked(bo); 243 ttm_mem_io_unlock(old_man); 244 } 245 246 /* 247 * Create and bind a ttm if required. 248 */ 249 250 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 251 if (bo->ttm == NULL) { 252 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 253 ret = ttm_tt_create(bo, zero); 254 if (ret) 255 goto out_err; 256 } 257 258 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 259 if (ret) 260 goto out_err; 261 262 if (mem->mem_type != TTM_PL_SYSTEM) { 263 ret = ttm_tt_bind(bo->ttm, mem, ctx); 264 if (ret) 265 goto out_err; 266 } 267 268 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 269 if (bdev->driver->move_notify) 270 bdev->driver->move_notify(bo, evict, mem); 271 bo->mem = *mem; 272 mem->mm_node = NULL; 273 goto moved; 274 } 275 } 276 277 if (bdev->driver->move_notify) 278 bdev->driver->move_notify(bo, evict, mem); 279 280 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 281 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 282 ret = ttm_bo_move_ttm(bo, ctx, mem); 283 else if (bdev->driver->move) 284 ret = bdev->driver->move(bo, evict, ctx, mem); 285 else 286 ret = ttm_bo_move_memcpy(bo, ctx, mem); 287 288 if (ret) { 289 if (bdev->driver->move_notify) { 290 swap(*mem, bo->mem); 291 bdev->driver->move_notify(bo, false, mem); 292 swap(*mem, bo->mem); 293 } 294 295 goto out_err; 296 } 297 298 moved: 299 if (bo->evicted) { 300 if (bdev->driver->invalidate_caches) { 301 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 302 if (ret) 303 pr_err("Can not flush read caches\n"); 304 } 305 bo->evicted = false; 306 } 307 308 if (bo->mem.mm_node) 309 bo->offset = (bo->mem.start << PAGE_SHIFT) + 310 bdev->man[bo->mem.mem_type].gpu_offset; 311 else 312 bo->offset = 0; 313 314 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT; 315 return 0; 316 317 out_err: 318 new_man = &bdev->man[bo->mem.mem_type]; 319 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 320 ttm_tt_destroy(bo->ttm); 321 bo->ttm = NULL; 322 } 323 324 return ret; 325 } 326 327 /** 328 * Call bo::reserved. 329 * Will release GPU memory type usage on destruction. 330 * This is the place to put in driver specific hooks to release 331 * driver private resources. 332 * Will release the bo::reserved lock. 333 */ 334 335 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 336 { 337 if (bo->bdev->driver->move_notify) 338 bo->bdev->driver->move_notify(bo, false, NULL); 339 340 ttm_tt_destroy(bo->ttm); 341 bo->ttm = NULL; 342 ttm_bo_mem_put(bo, &bo->mem); 343 } 344 345 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo) 346 { 347 int r; 348 349 if (bo->resv == &bo->ttm_resv) 350 return 0; 351 352 BUG_ON(!reservation_object_trylock(&bo->ttm_resv)); 353 354 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv); 355 if (r) 356 reservation_object_unlock(&bo->ttm_resv); 357 358 return r; 359 } 360 361 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo) 362 { 363 struct reservation_object_list *fobj; 364 struct dma_fence *fence; 365 int i; 366 367 fobj = reservation_object_get_list(&bo->ttm_resv); 368 fence = reservation_object_get_excl(&bo->ttm_resv); 369 if (fence && !fence->ops->signaled) 370 dma_fence_enable_sw_signaling(fence); 371 372 for (i = 0; fobj && i < fobj->shared_count; ++i) { 373 fence = rcu_dereference_protected(fobj->shared[i], 374 reservation_object_held(bo->resv)); 375 376 if (!fence->ops->signaled) 377 dma_fence_enable_sw_signaling(fence); 378 } 379 } 380 381 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 382 { 383 struct ttm_bo_device *bdev = bo->bdev; 384 struct ttm_bo_global *glob = bdev->glob; 385 int ret; 386 387 ret = ttm_bo_individualize_resv(bo); 388 if (ret) { 389 /* Last resort, if we fail to allocate memory for the 390 * fences block for the BO to become idle 391 */ 392 reservation_object_wait_timeout_rcu(bo->resv, true, false, 393 30 * HZ); 394 spin_lock(&glob->lru_lock); 395 goto error; 396 } 397 398 spin_lock(&glob->lru_lock); 399 ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY; 400 if (!ret) { 401 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) { 402 ttm_bo_del_from_lru(bo); 403 spin_unlock(&glob->lru_lock); 404 if (bo->resv != &bo->ttm_resv) 405 reservation_object_unlock(&bo->ttm_resv); 406 407 ttm_bo_cleanup_memtype_use(bo); 408 reservation_object_unlock(bo->resv); 409 return; 410 } 411 412 ttm_bo_flush_all_fences(bo); 413 414 /* 415 * Make NO_EVICT bos immediately available to 416 * shrinkers, now that they are queued for 417 * destruction. 418 */ 419 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 420 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 421 ttm_bo_add_to_lru(bo); 422 } 423 424 reservation_object_unlock(bo->resv); 425 } 426 if (bo->resv != &bo->ttm_resv) 427 reservation_object_unlock(&bo->ttm_resv); 428 429 error: 430 kref_get(&bo->list_kref); 431 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 432 spin_unlock(&glob->lru_lock); 433 434 schedule_delayed_work(&bdev->wq, 435 ((HZ / 100) < 1) ? 1 : HZ / 100); 436 } 437 438 /** 439 * function ttm_bo_cleanup_refs 440 * If bo idle, remove from delayed- and lru lists, and unref. 441 * If not idle, do nothing. 442 * 443 * Must be called with lru_lock and reservation held, this function 444 * will drop the lru lock and optionally the reservation lock before returning. 445 * 446 * @interruptible Any sleeps should occur interruptibly. 447 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 448 * @unlock_resv Unlock the reservation lock as well. 449 */ 450 451 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, 452 bool interruptible, bool no_wait_gpu, 453 bool unlock_resv) 454 { 455 struct ttm_bo_global *glob = bo->bdev->glob; 456 struct reservation_object *resv; 457 int ret; 458 459 if (unlikely(list_empty(&bo->ddestroy))) 460 resv = bo->resv; 461 else 462 resv = &bo->ttm_resv; 463 464 if (reservation_object_test_signaled_rcu(resv, true)) 465 ret = 0; 466 else 467 ret = -EBUSY; 468 469 if (ret && !no_wait_gpu) { 470 long lret; 471 472 if (unlock_resv) 473 reservation_object_unlock(bo->resv); 474 spin_unlock(&glob->lru_lock); 475 476 lret = reservation_object_wait_timeout_rcu(resv, true, 477 interruptible, 478 30 * HZ); 479 480 if (lret < 0) 481 return lret; 482 else if (lret == 0) 483 return -EBUSY; 484 485 spin_lock(&glob->lru_lock); 486 if (unlock_resv && !reservation_object_trylock(bo->resv)) { 487 /* 488 * We raced, and lost, someone else holds the reservation now, 489 * and is probably busy in ttm_bo_cleanup_memtype_use. 490 * 491 * Even if it's not the case, because we finished waiting any 492 * delayed destruction would succeed, so just return success 493 * here. 494 */ 495 spin_unlock(&glob->lru_lock); 496 return 0; 497 } 498 ret = 0; 499 } 500 501 if (ret || unlikely(list_empty(&bo->ddestroy))) { 502 if (unlock_resv) 503 reservation_object_unlock(bo->resv); 504 spin_unlock(&glob->lru_lock); 505 return ret; 506 } 507 508 ttm_bo_del_from_lru(bo); 509 list_del_init(&bo->ddestroy); 510 kref_put(&bo->list_kref, ttm_bo_ref_bug); 511 512 spin_unlock(&glob->lru_lock); 513 ttm_bo_cleanup_memtype_use(bo); 514 515 if (unlock_resv) 516 reservation_object_unlock(bo->resv); 517 518 return 0; 519 } 520 521 /** 522 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 523 * encountered buffers. 524 */ 525 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 526 { 527 struct ttm_bo_global *glob = bdev->glob; 528 struct list_head removed; 529 bool empty; 530 531 INIT_LIST_HEAD(&removed); 532 533 spin_lock(&glob->lru_lock); 534 while (!list_empty(&bdev->ddestroy)) { 535 struct ttm_buffer_object *bo; 536 537 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, 538 ddestroy); 539 kref_get(&bo->list_kref); 540 list_move_tail(&bo->ddestroy, &removed); 541 542 if (remove_all || bo->resv != &bo->ttm_resv) { 543 spin_unlock(&glob->lru_lock); 544 reservation_object_lock(bo->resv, NULL); 545 546 spin_lock(&glob->lru_lock); 547 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 548 549 } else if (reservation_object_trylock(bo->resv)) { 550 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 551 } else { 552 spin_unlock(&glob->lru_lock); 553 } 554 555 kref_put(&bo->list_kref, ttm_bo_release_list); 556 spin_lock(&glob->lru_lock); 557 } 558 list_splice_tail(&removed, &bdev->ddestroy); 559 empty = list_empty(&bdev->ddestroy); 560 spin_unlock(&glob->lru_lock); 561 562 return empty; 563 } 564 565 static void ttm_bo_delayed_workqueue(struct work_struct *work) 566 { 567 struct ttm_bo_device *bdev = 568 container_of(work, struct ttm_bo_device, wq.work); 569 570 if (!ttm_bo_delayed_delete(bdev, false)) 571 schedule_delayed_work(&bdev->wq, 572 ((HZ / 100) < 1) ? 1 : HZ / 100); 573 } 574 575 static void ttm_bo_release(struct kref *kref) 576 { 577 struct ttm_buffer_object *bo = 578 container_of(kref, struct ttm_buffer_object, kref); 579 struct ttm_bo_device *bdev = bo->bdev; 580 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 581 582 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 583 ttm_mem_io_lock(man, false); 584 ttm_mem_io_free_vm(bo); 585 ttm_mem_io_unlock(man); 586 ttm_bo_cleanup_refs_or_queue(bo); 587 kref_put(&bo->list_kref, ttm_bo_release_list); 588 } 589 590 void ttm_bo_put(struct ttm_buffer_object *bo) 591 { 592 kref_put(&bo->kref, ttm_bo_release); 593 } 594 EXPORT_SYMBOL(ttm_bo_put); 595 596 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 597 { 598 struct ttm_buffer_object *bo = *p_bo; 599 600 *p_bo = NULL; 601 ttm_bo_put(bo); 602 } 603 EXPORT_SYMBOL(ttm_bo_unref); 604 605 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 606 { 607 return cancel_delayed_work_sync(&bdev->wq); 608 } 609 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 610 611 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 612 { 613 if (resched) 614 schedule_delayed_work(&bdev->wq, 615 ((HZ / 100) < 1) ? 1 : HZ / 100); 616 } 617 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 618 619 static int ttm_bo_evict(struct ttm_buffer_object *bo, 620 struct ttm_operation_ctx *ctx) 621 { 622 struct ttm_bo_device *bdev = bo->bdev; 623 struct ttm_mem_reg evict_mem; 624 struct ttm_placement placement; 625 int ret = 0; 626 627 reservation_object_assert_held(bo->resv); 628 629 placement.num_placement = 0; 630 placement.num_busy_placement = 0; 631 bdev->driver->evict_flags(bo, &placement); 632 633 if (!placement.num_placement && !placement.num_busy_placement) { 634 ret = ttm_bo_pipeline_gutting(bo); 635 if (ret) 636 return ret; 637 638 return ttm_tt_create(bo, false); 639 } 640 641 evict_mem = bo->mem; 642 evict_mem.mm_node = NULL; 643 evict_mem.bus.io_reserved_vm = false; 644 evict_mem.bus.io_reserved_count = 0; 645 646 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx); 647 if (ret) { 648 if (ret != -ERESTARTSYS) { 649 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 650 bo); 651 ttm_bo_mem_space_debug(bo, &placement); 652 } 653 goto out; 654 } 655 656 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx); 657 if (unlikely(ret)) { 658 if (ret != -ERESTARTSYS) 659 pr_err("Buffer eviction failed\n"); 660 ttm_bo_mem_put(bo, &evict_mem); 661 goto out; 662 } 663 bo->evicted = true; 664 out: 665 return ret; 666 } 667 668 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 669 const struct ttm_place *place) 670 { 671 /* Don't evict this BO if it's outside of the 672 * requested placement range 673 */ 674 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 675 (place->lpfn && place->lpfn <= bo->mem.start)) 676 return false; 677 678 return true; 679 } 680 EXPORT_SYMBOL(ttm_bo_eviction_valuable); 681 682 /** 683 * Check the target bo is allowable to be evicted or swapout, including cases: 684 * 685 * a. if share same reservation object with ctx->resv, have assumption 686 * reservation objects should already be locked, so not lock again and 687 * return true directly when either the opreation allow_reserved_eviction 688 * or the target bo already is in delayed free list; 689 * 690 * b. Otherwise, trylock it. 691 */ 692 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo, 693 struct ttm_operation_ctx *ctx, bool *locked) 694 { 695 bool ret = false; 696 697 *locked = false; 698 if (bo->resv == ctx->resv) { 699 reservation_object_assert_held(bo->resv); 700 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT 701 || !list_empty(&bo->ddestroy)) 702 ret = true; 703 } else { 704 *locked = reservation_object_trylock(bo->resv); 705 ret = *locked; 706 } 707 708 return ret; 709 } 710 711 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 712 uint32_t mem_type, 713 const struct ttm_place *place, 714 struct ttm_operation_ctx *ctx) 715 { 716 struct ttm_bo_global *glob = bdev->glob; 717 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 718 struct ttm_buffer_object *bo = NULL; 719 bool locked = false; 720 unsigned i; 721 int ret; 722 723 spin_lock(&glob->lru_lock); 724 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 725 list_for_each_entry(bo, &man->lru[i], lru) { 726 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) 727 continue; 728 729 if (place && !bdev->driver->eviction_valuable(bo, 730 place)) { 731 if (locked) 732 reservation_object_unlock(bo->resv); 733 continue; 734 } 735 break; 736 } 737 738 /* If the inner loop terminated early, we have our candidate */ 739 if (&bo->lru != &man->lru[i]) 740 break; 741 742 bo = NULL; 743 } 744 745 if (!bo) { 746 spin_unlock(&glob->lru_lock); 747 return -EBUSY; 748 } 749 750 kref_get(&bo->list_kref); 751 752 if (!list_empty(&bo->ddestroy)) { 753 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible, 754 ctx->no_wait_gpu, locked); 755 kref_put(&bo->list_kref, ttm_bo_release_list); 756 return ret; 757 } 758 759 ttm_bo_del_from_lru(bo); 760 spin_unlock(&glob->lru_lock); 761 762 ret = ttm_bo_evict(bo, ctx); 763 if (locked) { 764 ttm_bo_unreserve(bo); 765 } else { 766 spin_lock(&glob->lru_lock); 767 ttm_bo_add_to_lru(bo); 768 spin_unlock(&glob->lru_lock); 769 } 770 771 kref_put(&bo->list_kref, ttm_bo_release_list); 772 return ret; 773 } 774 775 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 776 { 777 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 778 779 if (mem->mm_node) 780 (*man->func->put_node)(man, mem); 781 } 782 EXPORT_SYMBOL(ttm_bo_mem_put); 783 784 /** 785 * Add the last move fence to the BO and reserve a new shared slot. 786 */ 787 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 788 struct ttm_mem_type_manager *man, 789 struct ttm_mem_reg *mem) 790 { 791 struct dma_fence *fence; 792 int ret; 793 794 spin_lock(&man->move_lock); 795 fence = dma_fence_get(man->move); 796 spin_unlock(&man->move_lock); 797 798 if (fence) { 799 reservation_object_add_shared_fence(bo->resv, fence); 800 801 ret = reservation_object_reserve_shared(bo->resv); 802 if (unlikely(ret)) 803 return ret; 804 805 dma_fence_put(bo->moving); 806 bo->moving = fence; 807 } 808 809 return 0; 810 } 811 812 /** 813 * Repeatedly evict memory from the LRU for @mem_type until we create enough 814 * space, or we've evicted everything and there isn't enough space. 815 */ 816 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 817 uint32_t mem_type, 818 const struct ttm_place *place, 819 struct ttm_mem_reg *mem, 820 struct ttm_operation_ctx *ctx) 821 { 822 struct ttm_bo_device *bdev = bo->bdev; 823 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 824 int ret; 825 826 do { 827 ret = (*man->func->get_node)(man, bo, place, mem); 828 if (unlikely(ret != 0)) 829 return ret; 830 if (mem->mm_node) 831 break; 832 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx); 833 if (unlikely(ret != 0)) 834 return ret; 835 } while (1); 836 mem->mem_type = mem_type; 837 return ttm_bo_add_move_fence(bo, man, mem); 838 } 839 840 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 841 uint32_t cur_placement, 842 uint32_t proposed_placement) 843 { 844 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 845 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 846 847 /** 848 * Keep current caching if possible. 849 */ 850 851 if ((cur_placement & caching) != 0) 852 result |= (cur_placement & caching); 853 else if ((man->default_caching & caching) != 0) 854 result |= man->default_caching; 855 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 856 result |= TTM_PL_FLAG_CACHED; 857 else if ((TTM_PL_FLAG_WC & caching) != 0) 858 result |= TTM_PL_FLAG_WC; 859 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 860 result |= TTM_PL_FLAG_UNCACHED; 861 862 return result; 863 } 864 865 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 866 uint32_t mem_type, 867 const struct ttm_place *place, 868 uint32_t *masked_placement) 869 { 870 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 871 872 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 873 return false; 874 875 if ((place->flags & man->available_caching) == 0) 876 return false; 877 878 cur_flags |= (place->flags & man->available_caching); 879 880 *masked_placement = cur_flags; 881 return true; 882 } 883 884 /** 885 * Creates space for memory region @mem according to its type. 886 * 887 * This function first searches for free space in compatible memory types in 888 * the priority order defined by the driver. If free space isn't found, then 889 * ttm_bo_mem_force_space is attempted in priority order to evict and find 890 * space. 891 */ 892 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 893 struct ttm_placement *placement, 894 struct ttm_mem_reg *mem, 895 struct ttm_operation_ctx *ctx) 896 { 897 struct ttm_bo_device *bdev = bo->bdev; 898 struct ttm_mem_type_manager *man; 899 uint32_t mem_type = TTM_PL_SYSTEM; 900 uint32_t cur_flags = 0; 901 bool type_found = false; 902 bool type_ok = false; 903 bool has_erestartsys = false; 904 int i, ret; 905 906 ret = reservation_object_reserve_shared(bo->resv); 907 if (unlikely(ret)) 908 return ret; 909 910 mem->mm_node = NULL; 911 for (i = 0; i < placement->num_placement; ++i) { 912 const struct ttm_place *place = &placement->placement[i]; 913 914 ret = ttm_mem_type_from_place(place, &mem_type); 915 if (ret) 916 return ret; 917 man = &bdev->man[mem_type]; 918 if (!man->has_type || !man->use_type) 919 continue; 920 921 type_ok = ttm_bo_mt_compatible(man, mem_type, place, 922 &cur_flags); 923 924 if (!type_ok) 925 continue; 926 927 type_found = true; 928 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 929 cur_flags); 930 /* 931 * Use the access and other non-mapping-related flag bits from 932 * the memory placement flags to the current flags 933 */ 934 ttm_flag_masked(&cur_flags, place->flags, 935 ~TTM_PL_MASK_MEMTYPE); 936 937 if (mem_type == TTM_PL_SYSTEM) 938 break; 939 940 ret = (*man->func->get_node)(man, bo, place, mem); 941 if (unlikely(ret)) 942 return ret; 943 944 if (mem->mm_node) { 945 ret = ttm_bo_add_move_fence(bo, man, mem); 946 if (unlikely(ret)) { 947 (*man->func->put_node)(man, mem); 948 return ret; 949 } 950 break; 951 } 952 } 953 954 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 955 mem->mem_type = mem_type; 956 mem->placement = cur_flags; 957 return 0; 958 } 959 960 for (i = 0; i < placement->num_busy_placement; ++i) { 961 const struct ttm_place *place = &placement->busy_placement[i]; 962 963 ret = ttm_mem_type_from_place(place, &mem_type); 964 if (ret) 965 return ret; 966 man = &bdev->man[mem_type]; 967 if (!man->has_type || !man->use_type) 968 continue; 969 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 970 continue; 971 972 type_found = true; 973 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 974 cur_flags); 975 /* 976 * Use the access and other non-mapping-related flag bits from 977 * the memory placement flags to the current flags 978 */ 979 ttm_flag_masked(&cur_flags, place->flags, 980 ~TTM_PL_MASK_MEMTYPE); 981 982 if (mem_type == TTM_PL_SYSTEM) { 983 mem->mem_type = mem_type; 984 mem->placement = cur_flags; 985 mem->mm_node = NULL; 986 return 0; 987 } 988 989 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx); 990 if (ret == 0 && mem->mm_node) { 991 mem->placement = cur_flags; 992 return 0; 993 } 994 if (ret == -ERESTARTSYS) 995 has_erestartsys = true; 996 } 997 998 if (!type_found) { 999 pr_err(TTM_PFX "No compatible memory type found\n"); 1000 return -EINVAL; 1001 } 1002 1003 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1004 } 1005 EXPORT_SYMBOL(ttm_bo_mem_space); 1006 1007 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1008 struct ttm_placement *placement, 1009 struct ttm_operation_ctx *ctx) 1010 { 1011 int ret = 0; 1012 struct ttm_mem_reg mem; 1013 1014 reservation_object_assert_held(bo->resv); 1015 1016 mem.num_pages = bo->num_pages; 1017 mem.size = mem.num_pages << PAGE_SHIFT; 1018 mem.page_alignment = bo->mem.page_alignment; 1019 mem.bus.io_reserved_vm = false; 1020 mem.bus.io_reserved_count = 0; 1021 /* 1022 * Determine where to move the buffer. 1023 */ 1024 ret = ttm_bo_mem_space(bo, placement, &mem, ctx); 1025 if (ret) 1026 goto out_unlock; 1027 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx); 1028 out_unlock: 1029 if (ret && mem.mm_node) 1030 ttm_bo_mem_put(bo, &mem); 1031 return ret; 1032 } 1033 1034 static bool ttm_bo_places_compat(const struct ttm_place *places, 1035 unsigned num_placement, 1036 struct ttm_mem_reg *mem, 1037 uint32_t *new_flags) 1038 { 1039 unsigned i; 1040 1041 for (i = 0; i < num_placement; i++) { 1042 const struct ttm_place *heap = &places[i]; 1043 1044 if (mem->mm_node && (mem->start < heap->fpfn || 1045 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1046 continue; 1047 1048 *new_flags = heap->flags; 1049 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1050 (*new_flags & mem->placement & TTM_PL_MASK_MEM) && 1051 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) || 1052 (mem->placement & TTM_PL_FLAG_CONTIGUOUS))) 1053 return true; 1054 } 1055 return false; 1056 } 1057 1058 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1059 struct ttm_mem_reg *mem, 1060 uint32_t *new_flags) 1061 { 1062 if (ttm_bo_places_compat(placement->placement, placement->num_placement, 1063 mem, new_flags)) 1064 return true; 1065 1066 if ((placement->busy_placement != placement->placement || 1067 placement->num_busy_placement > placement->num_placement) && 1068 ttm_bo_places_compat(placement->busy_placement, 1069 placement->num_busy_placement, 1070 mem, new_flags)) 1071 return true; 1072 1073 return false; 1074 } 1075 EXPORT_SYMBOL(ttm_bo_mem_compat); 1076 1077 int ttm_bo_validate(struct ttm_buffer_object *bo, 1078 struct ttm_placement *placement, 1079 struct ttm_operation_ctx *ctx) 1080 { 1081 int ret; 1082 uint32_t new_flags; 1083 1084 reservation_object_assert_held(bo->resv); 1085 /* 1086 * Check whether we need to move buffer. 1087 */ 1088 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1089 ret = ttm_bo_move_buffer(bo, placement, ctx); 1090 if (ret) 1091 return ret; 1092 } else { 1093 /* 1094 * Use the access and other non-mapping-related flag bits from 1095 * the compatible memory placement flags to the active flags 1096 */ 1097 ttm_flag_masked(&bo->mem.placement, new_flags, 1098 ~TTM_PL_MASK_MEMTYPE); 1099 } 1100 /* 1101 * We might need to add a TTM. 1102 */ 1103 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1104 ret = ttm_tt_create(bo, true); 1105 if (ret) 1106 return ret; 1107 } 1108 return 0; 1109 } 1110 EXPORT_SYMBOL(ttm_bo_validate); 1111 1112 int ttm_bo_init_reserved(struct ttm_bo_device *bdev, 1113 struct ttm_buffer_object *bo, 1114 unsigned long size, 1115 enum ttm_bo_type type, 1116 struct ttm_placement *placement, 1117 uint32_t page_alignment, 1118 struct ttm_operation_ctx *ctx, 1119 size_t acc_size, 1120 struct sg_table *sg, 1121 struct reservation_object *resv, 1122 void (*destroy) (struct ttm_buffer_object *)) 1123 { 1124 int ret = 0; 1125 unsigned long num_pages; 1126 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1127 bool locked; 1128 1129 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx); 1130 if (ret) { 1131 pr_err("Out of kernel memory\n"); 1132 if (destroy) 1133 (*destroy)(bo); 1134 else 1135 kfree(bo); 1136 return -ENOMEM; 1137 } 1138 1139 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1140 if (num_pages == 0) { 1141 pr_err("Illegal buffer object size\n"); 1142 if (destroy) 1143 (*destroy)(bo); 1144 else 1145 kfree(bo); 1146 ttm_mem_global_free(mem_glob, acc_size); 1147 return -EINVAL; 1148 } 1149 bo->destroy = destroy ? destroy : ttm_bo_default_destroy; 1150 1151 kref_init(&bo->kref); 1152 kref_init(&bo->list_kref); 1153 atomic_set(&bo->cpu_writers, 0); 1154 INIT_LIST_HEAD(&bo->lru); 1155 INIT_LIST_HEAD(&bo->ddestroy); 1156 INIT_LIST_HEAD(&bo->swap); 1157 INIT_LIST_HEAD(&bo->io_reserve_lru); 1158 mutex_init(&bo->wu_mutex); 1159 bo->bdev = bdev; 1160 bo->type = type; 1161 bo->num_pages = num_pages; 1162 bo->mem.size = num_pages << PAGE_SHIFT; 1163 bo->mem.mem_type = TTM_PL_SYSTEM; 1164 bo->mem.num_pages = bo->num_pages; 1165 bo->mem.mm_node = NULL; 1166 bo->mem.page_alignment = page_alignment; 1167 bo->mem.bus.io_reserved_vm = false; 1168 bo->mem.bus.io_reserved_count = 0; 1169 bo->moving = NULL; 1170 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1171 bo->acc_size = acc_size; 1172 bo->sg = sg; 1173 if (resv) { 1174 bo->resv = resv; 1175 reservation_object_assert_held(bo->resv); 1176 } else { 1177 bo->resv = &bo->ttm_resv; 1178 } 1179 reservation_object_init(&bo->ttm_resv); 1180 atomic_inc(&bo->bdev->glob->bo_count); 1181 drm_vma_node_reset(&bo->vma_node); 1182 1183 /* 1184 * For ttm_bo_type_device buffers, allocate 1185 * address space from the device. 1186 */ 1187 if (bo->type == ttm_bo_type_device || 1188 bo->type == ttm_bo_type_sg) 1189 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1190 bo->mem.num_pages); 1191 1192 /* passed reservation objects should already be locked, 1193 * since otherwise lockdep will be angered in radeon. 1194 */ 1195 if (!resv) { 1196 locked = reservation_object_trylock(bo->resv); 1197 WARN_ON(!locked); 1198 } 1199 1200 if (likely(!ret)) 1201 ret = ttm_bo_validate(bo, placement, ctx); 1202 1203 if (unlikely(ret)) { 1204 if (!resv) 1205 ttm_bo_unreserve(bo); 1206 1207 ttm_bo_put(bo); 1208 return ret; 1209 } 1210 1211 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1212 spin_lock(&bdev->glob->lru_lock); 1213 ttm_bo_add_to_lru(bo); 1214 spin_unlock(&bdev->glob->lru_lock); 1215 } 1216 1217 return ret; 1218 } 1219 EXPORT_SYMBOL(ttm_bo_init_reserved); 1220 1221 int ttm_bo_init(struct ttm_bo_device *bdev, 1222 struct ttm_buffer_object *bo, 1223 unsigned long size, 1224 enum ttm_bo_type type, 1225 struct ttm_placement *placement, 1226 uint32_t page_alignment, 1227 bool interruptible, 1228 size_t acc_size, 1229 struct sg_table *sg, 1230 struct reservation_object *resv, 1231 void (*destroy) (struct ttm_buffer_object *)) 1232 { 1233 struct ttm_operation_ctx ctx = { interruptible, false }; 1234 int ret; 1235 1236 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement, 1237 page_alignment, &ctx, acc_size, 1238 sg, resv, destroy); 1239 if (ret) 1240 return ret; 1241 1242 if (!resv) 1243 ttm_bo_unreserve(bo); 1244 1245 return 0; 1246 } 1247 EXPORT_SYMBOL(ttm_bo_init); 1248 1249 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1250 unsigned long bo_size, 1251 unsigned struct_size) 1252 { 1253 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1254 size_t size = 0; 1255 1256 size += ttm_round_pot(struct_size); 1257 size += ttm_round_pot(npages * sizeof(void *)); 1258 size += ttm_round_pot(sizeof(struct ttm_tt)); 1259 return size; 1260 } 1261 EXPORT_SYMBOL(ttm_bo_acc_size); 1262 1263 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1264 unsigned long bo_size, 1265 unsigned struct_size) 1266 { 1267 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1268 size_t size = 0; 1269 1270 size += ttm_round_pot(struct_size); 1271 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1272 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1273 return size; 1274 } 1275 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1276 1277 int ttm_bo_create(struct ttm_bo_device *bdev, 1278 unsigned long size, 1279 enum ttm_bo_type type, 1280 struct ttm_placement *placement, 1281 uint32_t page_alignment, 1282 bool interruptible, 1283 struct ttm_buffer_object **p_bo) 1284 { 1285 struct ttm_buffer_object *bo; 1286 size_t acc_size; 1287 int ret; 1288 1289 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1290 if (unlikely(bo == NULL)) 1291 return -ENOMEM; 1292 1293 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1294 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1295 interruptible, acc_size, 1296 NULL, NULL, NULL); 1297 if (likely(ret == 0)) 1298 *p_bo = bo; 1299 1300 return ret; 1301 } 1302 EXPORT_SYMBOL(ttm_bo_create); 1303 1304 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1305 unsigned mem_type) 1306 { 1307 struct ttm_operation_ctx ctx = { 1308 .interruptible = false, 1309 .no_wait_gpu = false, 1310 .flags = TTM_OPT_FLAG_FORCE_ALLOC 1311 }; 1312 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1313 struct ttm_bo_global *glob = bdev->glob; 1314 struct dma_fence *fence; 1315 int ret; 1316 unsigned i; 1317 1318 /* 1319 * Can't use standard list traversal since we're unlocking. 1320 */ 1321 1322 spin_lock(&glob->lru_lock); 1323 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1324 while (!list_empty(&man->lru[i])) { 1325 spin_unlock(&glob->lru_lock); 1326 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx); 1327 if (ret) 1328 return ret; 1329 spin_lock(&glob->lru_lock); 1330 } 1331 } 1332 spin_unlock(&glob->lru_lock); 1333 1334 spin_lock(&man->move_lock); 1335 fence = dma_fence_get(man->move); 1336 spin_unlock(&man->move_lock); 1337 1338 if (fence) { 1339 ret = dma_fence_wait(fence, false); 1340 dma_fence_put(fence); 1341 if (ret) 1342 return ret; 1343 } 1344 1345 return 0; 1346 } 1347 1348 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1349 { 1350 struct ttm_mem_type_manager *man; 1351 int ret = -EINVAL; 1352 1353 if (mem_type >= TTM_NUM_MEM_TYPES) { 1354 pr_err("Illegal memory type %d\n", mem_type); 1355 return ret; 1356 } 1357 man = &bdev->man[mem_type]; 1358 1359 if (!man->has_type) { 1360 pr_err("Trying to take down uninitialized memory manager type %u\n", 1361 mem_type); 1362 return ret; 1363 } 1364 1365 man->use_type = false; 1366 man->has_type = false; 1367 1368 ret = 0; 1369 if (mem_type > 0) { 1370 ret = ttm_bo_force_list_clean(bdev, mem_type); 1371 if (ret) { 1372 pr_err("Cleanup eviction failed\n"); 1373 return ret; 1374 } 1375 1376 ret = (*man->func->takedown)(man); 1377 } 1378 1379 dma_fence_put(man->move); 1380 man->move = NULL; 1381 1382 return ret; 1383 } 1384 EXPORT_SYMBOL(ttm_bo_clean_mm); 1385 1386 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1387 { 1388 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1389 1390 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1391 pr_err("Illegal memory manager memory type %u\n", mem_type); 1392 return -EINVAL; 1393 } 1394 1395 if (!man->has_type) { 1396 pr_err("Memory type %u has not been initialized\n", mem_type); 1397 return 0; 1398 } 1399 1400 return ttm_bo_force_list_clean(bdev, mem_type); 1401 } 1402 EXPORT_SYMBOL(ttm_bo_evict_mm); 1403 1404 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1405 unsigned long p_size) 1406 { 1407 int ret; 1408 struct ttm_mem_type_manager *man; 1409 unsigned i; 1410 1411 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1412 man = &bdev->man[type]; 1413 BUG_ON(man->has_type); 1414 man->io_reserve_fastpath = true; 1415 man->use_io_reserve_lru = false; 1416 mutex_init(&man->io_reserve_mutex); 1417 spin_lock_init(&man->move_lock); 1418 INIT_LIST_HEAD(&man->io_reserve_lru); 1419 1420 ret = bdev->driver->init_mem_type(bdev, type, man); 1421 if (ret) 1422 return ret; 1423 man->bdev = bdev; 1424 1425 if (type != TTM_PL_SYSTEM) { 1426 ret = (*man->func->init)(man, p_size); 1427 if (ret) 1428 return ret; 1429 } 1430 man->has_type = true; 1431 man->use_type = true; 1432 man->size = p_size; 1433 1434 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1435 INIT_LIST_HEAD(&man->lru[i]); 1436 man->move = NULL; 1437 1438 return 0; 1439 } 1440 EXPORT_SYMBOL(ttm_bo_init_mm); 1441 1442 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1443 { 1444 struct ttm_bo_global *glob = 1445 container_of(kobj, struct ttm_bo_global, kobj); 1446 1447 __free_page(glob->dummy_read_page); 1448 kfree(glob); 1449 } 1450 1451 void ttm_bo_global_release(struct drm_global_reference *ref) 1452 { 1453 struct ttm_bo_global *glob = ref->object; 1454 1455 kobject_del(&glob->kobj); 1456 kobject_put(&glob->kobj); 1457 } 1458 EXPORT_SYMBOL(ttm_bo_global_release); 1459 1460 int ttm_bo_global_init(struct drm_global_reference *ref) 1461 { 1462 struct ttm_bo_global_ref *bo_ref = 1463 container_of(ref, struct ttm_bo_global_ref, ref); 1464 struct ttm_bo_global *glob = ref->object; 1465 int ret; 1466 unsigned i; 1467 1468 mutex_init(&glob->device_list_mutex); 1469 spin_lock_init(&glob->lru_lock); 1470 glob->mem_glob = bo_ref->mem_glob; 1471 glob->mem_glob->bo_glob = glob; 1472 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1473 1474 if (unlikely(glob->dummy_read_page == NULL)) { 1475 ret = -ENOMEM; 1476 goto out_no_drp; 1477 } 1478 1479 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1480 INIT_LIST_HEAD(&glob->swap_lru[i]); 1481 INIT_LIST_HEAD(&glob->device_list); 1482 atomic_set(&glob->bo_count, 0); 1483 1484 ret = kobject_init_and_add( 1485 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1486 if (unlikely(ret != 0)) 1487 kobject_put(&glob->kobj); 1488 return ret; 1489 out_no_drp: 1490 kfree(glob); 1491 return ret; 1492 } 1493 EXPORT_SYMBOL(ttm_bo_global_init); 1494 1495 1496 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1497 { 1498 int ret = 0; 1499 unsigned i = TTM_NUM_MEM_TYPES; 1500 struct ttm_mem_type_manager *man; 1501 struct ttm_bo_global *glob = bdev->glob; 1502 1503 while (i--) { 1504 man = &bdev->man[i]; 1505 if (man->has_type) { 1506 man->use_type = false; 1507 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1508 ret = -EBUSY; 1509 pr_err("DRM memory manager type %d is not clean\n", 1510 i); 1511 } 1512 man->has_type = false; 1513 } 1514 } 1515 1516 mutex_lock(&glob->device_list_mutex); 1517 list_del(&bdev->device_list); 1518 mutex_unlock(&glob->device_list_mutex); 1519 1520 cancel_delayed_work_sync(&bdev->wq); 1521 1522 if (ttm_bo_delayed_delete(bdev, true)) 1523 pr_debug("Delayed destroy list was clean\n"); 1524 1525 spin_lock(&glob->lru_lock); 1526 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1527 if (list_empty(&bdev->man[0].lru[0])) 1528 pr_debug("Swap list %d was clean\n", i); 1529 spin_unlock(&glob->lru_lock); 1530 1531 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1532 1533 return ret; 1534 } 1535 EXPORT_SYMBOL(ttm_bo_device_release); 1536 1537 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1538 struct ttm_bo_global *glob, 1539 struct ttm_bo_driver *driver, 1540 struct address_space *mapping, 1541 uint64_t file_page_offset, 1542 bool need_dma32) 1543 { 1544 int ret = -EINVAL; 1545 1546 bdev->driver = driver; 1547 1548 memset(bdev->man, 0, sizeof(bdev->man)); 1549 1550 /* 1551 * Initialize the system memory buffer type. 1552 * Other types need to be driver / IOCTL initialized. 1553 */ 1554 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1555 if (unlikely(ret != 0)) 1556 goto out_no_sys; 1557 1558 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset, 1559 0x10000000); 1560 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1561 INIT_LIST_HEAD(&bdev->ddestroy); 1562 bdev->dev_mapping = mapping; 1563 bdev->glob = glob; 1564 bdev->need_dma32 = need_dma32; 1565 mutex_lock(&glob->device_list_mutex); 1566 list_add_tail(&bdev->device_list, &glob->device_list); 1567 mutex_unlock(&glob->device_list_mutex); 1568 1569 return 0; 1570 out_no_sys: 1571 return ret; 1572 } 1573 EXPORT_SYMBOL(ttm_bo_device_init); 1574 1575 /* 1576 * buffer object vm functions. 1577 */ 1578 1579 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1580 { 1581 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1582 1583 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1584 if (mem->mem_type == TTM_PL_SYSTEM) 1585 return false; 1586 1587 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1588 return false; 1589 1590 if (mem->placement & TTM_PL_FLAG_CACHED) 1591 return false; 1592 } 1593 return true; 1594 } 1595 1596 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1597 { 1598 struct ttm_bo_device *bdev = bo->bdev; 1599 1600 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping); 1601 ttm_mem_io_free_vm(bo); 1602 } 1603 1604 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1605 { 1606 struct ttm_bo_device *bdev = bo->bdev; 1607 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1608 1609 ttm_mem_io_lock(man, false); 1610 ttm_bo_unmap_virtual_locked(bo); 1611 ttm_mem_io_unlock(man); 1612 } 1613 1614 1615 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1616 1617 int ttm_bo_wait(struct ttm_buffer_object *bo, 1618 bool interruptible, bool no_wait) 1619 { 1620 long timeout = 15 * HZ; 1621 1622 if (no_wait) { 1623 if (reservation_object_test_signaled_rcu(bo->resv, true)) 1624 return 0; 1625 else 1626 return -EBUSY; 1627 } 1628 1629 timeout = reservation_object_wait_timeout_rcu(bo->resv, true, 1630 interruptible, timeout); 1631 if (timeout < 0) 1632 return timeout; 1633 1634 if (timeout == 0) 1635 return -EBUSY; 1636 1637 reservation_object_add_excl_fence(bo->resv, NULL); 1638 return 0; 1639 } 1640 EXPORT_SYMBOL(ttm_bo_wait); 1641 1642 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1643 { 1644 int ret = 0; 1645 1646 /* 1647 * Using ttm_bo_reserve makes sure the lru lists are updated. 1648 */ 1649 1650 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1651 if (unlikely(ret != 0)) 1652 return ret; 1653 ret = ttm_bo_wait(bo, true, no_wait); 1654 if (likely(ret == 0)) 1655 atomic_inc(&bo->cpu_writers); 1656 ttm_bo_unreserve(bo); 1657 return ret; 1658 } 1659 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1660 1661 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1662 { 1663 atomic_dec(&bo->cpu_writers); 1664 } 1665 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1666 1667 /** 1668 * A buffer object shrink method that tries to swap out the first 1669 * buffer object on the bo_global::swap_lru list. 1670 */ 1671 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx) 1672 { 1673 struct ttm_buffer_object *bo; 1674 int ret = -EBUSY; 1675 bool locked; 1676 unsigned i; 1677 1678 spin_lock(&glob->lru_lock); 1679 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1680 list_for_each_entry(bo, &glob->swap_lru[i], swap) { 1681 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) { 1682 ret = 0; 1683 break; 1684 } 1685 } 1686 if (!ret) 1687 break; 1688 } 1689 1690 if (ret) { 1691 spin_unlock(&glob->lru_lock); 1692 return ret; 1693 } 1694 1695 kref_get(&bo->list_kref); 1696 1697 if (!list_empty(&bo->ddestroy)) { 1698 ret = ttm_bo_cleanup_refs(bo, false, false, locked); 1699 kref_put(&bo->list_kref, ttm_bo_release_list); 1700 return ret; 1701 } 1702 1703 ttm_bo_del_from_lru(bo); 1704 spin_unlock(&glob->lru_lock); 1705 1706 /** 1707 * Move to system cached 1708 */ 1709 1710 if (bo->mem.mem_type != TTM_PL_SYSTEM || 1711 bo->ttm->caching_state != tt_cached) { 1712 struct ttm_operation_ctx ctx = { false, false }; 1713 struct ttm_mem_reg evict_mem; 1714 1715 evict_mem = bo->mem; 1716 evict_mem.mm_node = NULL; 1717 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1718 evict_mem.mem_type = TTM_PL_SYSTEM; 1719 1720 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx); 1721 if (unlikely(ret != 0)) 1722 goto out; 1723 } 1724 1725 /** 1726 * Make sure BO is idle. 1727 */ 1728 1729 ret = ttm_bo_wait(bo, false, false); 1730 if (unlikely(ret != 0)) 1731 goto out; 1732 1733 ttm_bo_unmap_virtual(bo); 1734 1735 /** 1736 * Swap out. Buffer will be swapped in again as soon as 1737 * anyone tries to access a ttm page. 1738 */ 1739 1740 if (bo->bdev->driver->swap_notify) 1741 bo->bdev->driver->swap_notify(bo); 1742 1743 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1744 out: 1745 1746 /** 1747 * 1748 * Unreserve without putting on LRU to avoid swapping out an 1749 * already swapped buffer. 1750 */ 1751 if (locked) 1752 reservation_object_unlock(bo->resv); 1753 kref_put(&bo->list_kref, ttm_bo_release_list); 1754 return ret; 1755 } 1756 EXPORT_SYMBOL(ttm_bo_swapout); 1757 1758 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1759 { 1760 struct ttm_operation_ctx ctx = { 1761 .interruptible = false, 1762 .no_wait_gpu = false 1763 }; 1764 1765 while (ttm_bo_swapout(bdev->glob, &ctx) == 0) 1766 ; 1767 } 1768 EXPORT_SYMBOL(ttm_bo_swapout_all); 1769 1770 /** 1771 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1772 * unreserved 1773 * 1774 * @bo: Pointer to buffer 1775 */ 1776 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1777 { 1778 int ret; 1779 1780 /* 1781 * In the absense of a wait_unlocked API, 1782 * Use the bo::wu_mutex to avoid triggering livelocks due to 1783 * concurrent use of this function. Note that this use of 1784 * bo::wu_mutex can go away if we change locking order to 1785 * mmap_sem -> bo::reserve. 1786 */ 1787 ret = mutex_lock_interruptible(&bo->wu_mutex); 1788 if (unlikely(ret != 0)) 1789 return -ERESTARTSYS; 1790 if (!ww_mutex_is_locked(&bo->resv->lock)) 1791 goto out_unlock; 1792 ret = reservation_object_lock_interruptible(bo->resv, NULL); 1793 if (ret == -EINTR) 1794 ret = -ERESTARTSYS; 1795 if (unlikely(ret != 0)) 1796 goto out_unlock; 1797 reservation_object_unlock(bo->resv); 1798 1799 out_unlock: 1800 mutex_unlock(&bo->wu_mutex); 1801 return ret; 1802 } 1803